| Recurring |
multiple_organization |
(a) In the article, it is mentioned that Rank One, a facial recognition provider, experienced an increase in error rates when face masks were added to the recognition process. The CEO of Rank One, Brendan Klare, stated that the company's algorithm had an error rate of 0.6% without masks, which increased to 34.5% once masks were digitally applied. Additionally, the CEO mentioned that the company had started offering "periocular recognition" to identify people just off their eyes and nose. However, Rank One was not able to submit this algorithm to NIST due to submission limits [103647].
(b) The article also discusses Trueface, another facial recognition company used in schools and on Air Force bases, which saw its algorithm error rate increase from 0.9% to 34.8% when masks were added. The CEO of Trueface, Shaun Moore, mentioned that the company was working on a better algorithm for detecting beyond masks and had made improvements in their technology to identify individuals wearing masks. Moore stated that since their latest submission to NIST, they had trained their algorithm with more mask data and deployed the improved technology in the real world [103647]. |
| Phase (Design/Operation) |
design, operation |
(a) The article discusses how various facial recognition algorithms designed with face masks in mind claimed they could accurately identify people even with half of their face covered. However, the latest results from the US National Institute of Standards and Technology show that these algorithms experienced marginal increases in error rates once masks were introduced. For example, Rank One's algorithm had an error rate of 0.6% without masks, but this increased to 34.5% once masks were digitally applied. This indicates a failure in the design phase where the algorithms did not perform as expected when faced with the real-world scenario of people wearing masks [103647].
(b) The article also mentions how the software failure incident related to the operation phase, specifically mentioning the company Rank One. Rank One's facial recognition algorithm had an error rate of 0.6% without masks, but this increased to 34.5% once masks were digitally applied. This indicates a failure in the operation phase where the algorithm's performance was impacted by the introduction of masks, potentially due to limitations in the operation or misuse of the system in handling masked faces [103647]. |
| Boundary (Internal/External) |
within_system |
(a) within_system: The software failure incident discussed in the articles is related to facial recognition algorithms struggling to accurately identify individuals when they are wearing face masks. The error rates of these algorithms increased significantly when masks were added, indicating a failure within the system in terms of adapting to this new challenge [103647]. The algorithms were designed with face masks in mind but still experienced higher error rates, suggesting limitations within the software itself in handling this specific scenario. |
| Nature (Human/Non-human) |
non-human_actions, human_actions |
(a) The software failure incident occurring due to non-human actions:
The software failure in this case is related to the inability of facial recognition algorithms to accurately identify individuals wearing face masks. The error rates increased significantly when masks were added to the images, indicating a failure of the algorithms to perform effectively in the presence of this new factor introduced by the non-human action of wearing masks [103647].
(b) The software failure incident occurring due to human actions:
In this case, the failure of the facial recognition algorithms to accurately identify individuals wearing masks can also be attributed to human actions. The decision by individuals to wear face masks as a preventive measure against the spread of the novel coronavirus introduced a new challenge for the algorithms, leading to increased error rates. Additionally, the human action of developing and training the algorithms without considering face masks as a factor contributed to the failure [103647]. |
| Dimension (Hardware/Software) |
hardware, software |
(a) The software failure incident related to hardware:
- The article discusses the impact of face masks on facial recognition algorithms, which are software-based systems. The errors in recognition accuracy are attributed to the physical presence of masks, which can be considered a hardware-related factor affecting the software's performance [103647].
(b) The software failure incident related to software:
- The main focus of the article is on the software failure incident in facial recognition algorithms due to the introduction of face masks. The errors in identification accuracy are directly linked to the software's inability to adapt to the new environmental factor of face masks, indicating a software-related failure [103647]. |
| Objective (Malicious/Non-malicious) |
non-malicious |
(a) The software failure incident discussed in the article is non-malicious. The failure is related to the inability of facial recognition algorithms to accurately identify individuals when they are wearing face masks. This issue arose due to the introduction of face masks as a factor affecting the algorithms' performance, rather than any malicious intent by humans to harm the system [103647]. |
| Intent (Poor/Accidental Decisions) |
poor_decisions, accidental_decisions |
(a) The intent of the software failure incident related to poor decisions can be seen in the article. For example, Rank One CEO Brendan Klare mentioned that the company wasn't able to submit their algorithm designed for identifying people with masks to NIST due to the agency's limit to one submission per organization. This limitation could be considered a poor decision that hindered the company's ability to showcase their improved algorithm [103647].
(b) The intent of the software failure incident related to accidental decisions can be observed in the article as well. Trueface CEO Shaun Moore mentioned that their algorithm graded by NIST was not trained with face masks in mind, indicating an unintentional decision in the algorithm's development process. However, the company was making improvements to address this issue [103647]. |
| Capability (Incompetence/Accidental) |
development_incompetence |
(a) In the article, it is mentioned that some facial recognition algorithms experienced significant increases in error rates once face masks were introduced. For example, Rank One's algorithm had an error rate of 0.6% without masks, but this increased to 34.5% once masks were digitally applied. This indicates a potential failure due to development incompetence, as the algorithms may not have been adequately designed or trained to handle the new challenge of identifying individuals with face masks [103647].
(b) The article also highlights that some companies, like Trueface, saw their algorithm error rates increase from 0.9% to 34.8% once masks were added. Trueface's CEO mentioned that their algorithm was not initially trained with face masks in mind, indicating a potential accidental failure where the introduction of face masks was not anticipated or properly addressed during the development phase [103647]. |
| Duration |
temporary |
The software failure incident discussed in the articles is more likely to be temporary rather than permanent. This is evident from the fact that the error rates in facial recognition algorithms increased when face masks were introduced, indicating that the failure was due to specific circumstances (introduction of face masks) rather than a fundamental flaw in the algorithms themselves [103647]. Additionally, the article mentions that companies like Trueface and Rank One are actively working on improving their algorithms to better detect individuals wearing masks, showing that the issue is being addressed and can potentially be resolved [103647]. |
| Behaviour |
crash, omission, value, other |
(a) crash: The software failure incident related to facial recognition algorithms can be categorized as a crash. The algorithms experienced marginal increases in error rates once masks were introduced, with some algorithms having error rates increasing up to 99% [103647].
(b) omission: The software failure incident can also be related to omission. For example, Rank One's facial recognition algorithm had an error rate of 0.6% without masks, but this error rate increased to 34.5% once masks were digitally applied, indicating an omission in accurately identifying individuals when masks were present [103647].
(c) timing: The timing of the software failure incident is not explicitly mentioned in the article.
(d) value: The software failure incident can be related to a failure in value. For instance, Trueface's algorithm error rate increased from 0.9% to 34.8% once masks were added, indicating a failure in correctly identifying individuals wearing masks [103647].
(e) byzantine: The software failure incident is not related to a byzantine behavior as there is no mention of inconsistent responses or interactions in the article.
(f) other: The other behavior exhibited by the software failure incident is the inability of the algorithms to accurately identify individuals when their faces are covered with masks, leading to increased error rates. This behavior can be attributed to the challenges faced by facial recognition companies in adapting their algorithms to work effectively in the presence of face masks [103647]. |